[bfrancojr]

Fixes #176.

cpp now keeps a comment that no entry follows, as protowire-go does since protowire-go#283 (Go's #279). Without this, the differential against Go main fails on every PR, the release PR #175 included.

What changes for a reader of the AST

• A comment at the end of a document goes to the new Document::trailing_comments. It was dropped.
• A comment before a block's } goes to the new Block::trailing_comments or BlockVal::trailing_comments. It was dropped, or moved out of the block onto the next entry.
• Comments before a dataset row's (, and inside the row, go to the new DatasetRow::leading_comments. They went to the next directive, or were dropped. Parse and DatasetReader set them the same way. The reader now parses a row from the start of its buffer, as Go's parseRow does, and the parser mode that dropped those comments is gone.
• FormatDocument writes each where it was. A comment inside a row comes out before the row, since a row is written on one line. CanonicalizeKeyed keeps the comments through the blocks it builds, and the Rewriter's copies keep them too.
• Comments inside a list still move after it, as in Go (protowire-go#282).

Evidence

• scripts/go_differential.sh with spec dd8e221 and Go b3c210f: before this change, ParseTolerant differs (2 lines, both the # chords comment in testdata/differential/rewrite/settings.pxf); after it, no differences. The dumpers on both sides (cmd/pxf_diff, tools/godiff) now print the new fields, and Go's output is one line longer than before, which is that comment.
• New test/pxf_closing_comments_test.cc (4 tests) and Keyed.CanonicalizeKeepsClosingComments port Go's format_comments_test.go with the same documents and expected output. All 5 fail with the parser, formatter, keyed and reader changes reverted, and pass with them.
• ctest passes 747 of 747 locally. clang-format 18 has been applied. Local clang-tidy 23 reports nothing on the 5 changed source files.
• Speed: streaming a 300,000-row table, best of 5, run twice. Plain rows take 84.5 ms against main's 81 ms (+4%). Rows that each carry a trailing comment take 91.5 ms against 84 ms (+9%), since those comments are now kept rather than skipped.
• CI on f82e2fc (head): all 11 checks pass, including the differential against protowire-go (now main b3c210f), CodeQL, clang-tidy 18, ASan + UBSan and Windows.

Not ported

Go's new FuzzFormatDocumentFixedPoint fuzzer is not ported: this repo has no fuzz targets at all (no LLVMFuzzerTestOneInput anywhere). Every new test does check the format → parse → format fixed point.

Then

Once this merges, I'll rebase the release PR #175 so 2.1.0 includes it.